Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/321806603

Best-MatchedInternalStandardNormalizationinLiquidChromatography-MassSpectrometryMetabolomicsAppliedto....

ArticleinAnalyticalChemistry·December2017

DOI:10.1021/acs.analchem.7b04400

CITATION

1

READS

47

4authors,including:

KatherineR.Heal

UniversityofWashingtonSeattle

60PUBLICATIONS165CITATIONS

SEEPROFILE

AnitraEIngalls

UniversityofWashingtonSeattle

118PUBLICATIONS2,435CITATIONS

SEEPROFILE

AllcontentfollowingthispagewasuploadedbyKatherineR.Healon03January2018.

Theuserhasrequestedenhancementofthedownloadedfile.

Best-Matched Internal Standard Normalization in LiquidChromatography−Mass Spectrometry Metabolomics Applied toEnvironmental SamplesAngela K. Boysen,†,‡ Katherine R. Heal,†,‡ Laura T. Carlson,† and Anitra E. Ingalls*,†

†School of Oceanography, University of Washington, Seattle, Washington, United States

*S Supporting Information

ABSTRACT: The goal of metabolomics is to measure theentire range of small organic molecules in biological samples.In liquid chromatography−mass spectrometry-based metab-olomics, formidable analytical challenges remain in removingthe nonbiological factors that affect chromatographic peakareas. These factors include sample matrix-induced ionsuppression, chromatographic quality, and analytical drift.The combination of these factors is referred to as obscuringvariation. Some metabolomics samples can exhibit intenseobscuring variation due to matrix-induced ion suppression,rendering large amounts of data unreliable and difficult tointerpret. Existing normalization techniques have limitedapplicability to these sample types. Here we present a datanormalization method to minimize the effects of obscuringvariation. We normalize peak areas using a batch-specific normalization process, which matches measured metabolites withisotope-labeled internal standards that behave similarly during the analysis. This method, called best-matched internal standard(B-MIS) normalization, can be applied to targeted or untargeted metabolomics data sets and yields relative concentrations. Weevaluate and demonstrate the utility of B-MIS normalization using marine environmental samples and laboratory grown culturesof phytoplankton. In untargeted analyses, B-MIS normalization allowed for inclusion of mass features in downstream analysesthat would have been considered unreliable without normalization due to obscuring variation. B-MIS normalization for targetedor untargeted metabolomics is freely available at https://github.com/IngallsLabUW/B-MIS-normalization.

Liquid chromatography−mass spectrometry (LC-MS)-based metabolomics has emerged as an important

scientific tool over the past decade, with the potential toidentify and quantify thousands of compounds resulting fromcellular activity.1−3 Quantification relies on an analyte’sconcentration being proportional to its peak area. However,independent of analyte concentration, peak areas can beaffected by sample-matrix-induced ion suppression, injectionvolume, chromatographic quality, or analytical drift over time;the overall effects of these nonbiological factors are referred tohere as obscuring variation since they may mask biologicaldifferences of interest.4 Performing metabolomics analyses onenvironmental samples is especially challenging becausecomplex matrices such as natural seawater or culture mediaintroduce variability in ionization efficiency, but there are nouniversally accepted normalization techniques for minimizingobscuring variation in environmental samples. Here we presenta normalization method that addresses this outstandingchallenge in LC-MS-based metabolomics.In biomedical metabolomics, there are several approaches to

minimize obscuring variation that rely on pre- or post-acquisition normalization. However, application of theseapproaches to environmental metabolomics is limited, and

little attention has been given to removing obscuring variationin environmental samples.3,5−12 Preacquisition normalizationapproaches ensure that samples are at similar concentrationsbefore analysis to minimize the effect of matrix-induced ionsuppression. Postacquisition normalization techniques fall intotwo general categories, those that use a quality control sampleinjected several times over the course of a run to correct foranalytical drift, and those that employ isotope-labeled internalstandards.Analytical drift over time can be normalized postacquisition

by applying a quality control-based locally weighted scatterplotsmoothing (LOESS)13,14 or vector regression;15 theseapproaches attempt to minimize one aspect of obscuringvariation (analytical drift) and assume that the sample-matrixinduced ion suppression is consistent between samples. Thisapproach may be appropriate in biomedical studies whenprenormalization can be carefully applied (for instance, byadjusting sample concentration or injection volume so thatthere is a consistent amount of creatinine in each sample, as

Received: October 25, 2017Accepted: December 14, 2017Published: December 14, 2017

Article

pubs.acs.org/acCite This: Anal. Chem. XXXX, XXX, XXX−XXX

© XXXX American Chemical Society A DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

done routinely in urine metabolomics16,17). For complexmatrices, prenormalization is often impossible due to thevariability in matrix among samples or lack of informationabout sample size. For instance, if cell density is not knownuntil after sample collection, extracting a consistent amount ofbiomass is not possible. Furthermore, these approaches requirelarge sample sets and tens of injections of a pooled sample as aquality control, which is not always possible in metabolomicsstudies with limited sample size and few samples to compare.Other postacquisition strategies use isotope-labeled internal

standards to minimize obscuring variation. Some studiesnormalize each metabolite peak area to a single isotope-labeled internal standard.18 This approach is very sensitive tothe internal standard used for normalization and assumes thatall compounds experience the same obscuring variation as theinternal standard; these caveats make this approach limited inutility. Other studies use multiple internal standards, applyingeach to analytes that elute within a prescribed retention timewindow.19,20 Although this may be an improvement over asingle internal standard, these approaches do not consider thefact that obscuring variation is not fully described by retentiontime. Our group has previously used a hybrid approach thatmatches isotope-labeled internal standards to analytes forwhich commercially available isotope-labeled internal stand-

ards are unavailable,21,22 but this has not been applied to full-scale metabolomics analyses.In laboratory studies, growing an organism on a 13C-labeled

substrate can yield a valuable 13C-labeled stock used tocalculate accurate concentrations of both targeted anduntargeted metabolites and aid in metabolite identification,9,23

but this approach cannot be applied to studies of mixedcommunities in the environment. More sophisticated ap-proaches attempt to model the variability of each peak by acombination of internal standards.4,24 These methods requiretens of injections of a pooled sample to obtain sufficient datafor model training and make use of logarithmic trans-formations, which may limit quantitative comparison andhinder intuitive interpretation. To our knowledge, thesemethods have not been applied to targeted metabolomicanalyses.Given the paucity of normalization approaches that are

appropriate for minimizing obscuring variation in samples witha high degree of complexity in both analytes and matrices, wepresent a method called best-matched internal standard (B-MIS) normalization. This method has the following advantagesover existing techniques: (1) B-MIS normalization can beapplied to targeted and untargeted metabolomics studies withsample sets of any size or nature. (2) B-MIS normalization canbe customized to accommodate any number and variety of

Scheme 1. Overview of Data Acquisition and Processinga

aAfter extraction, aqueous fractions were run on both RP and HILIC, while organic fractions were run only on RP. These fractions were run onboth the TQS and QE. Skyline27 was used for integrations for targeted analyses. XCMS29−31 was used for peak-picking and integrations foruntargeted analyses. Data were subjected to quality control filtering as described in the text. For each compound detected, the B-MIS was chosenvia the steps shown. RSD = relative standard deviation. *Only applicable in targeted analyses. **This cut-off is user defined; in our analyses, aninternal standard was considered acceptable if it decreased the RSD of the pooled injections by 40%.

Analytical Chemistry Article

DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

B

internal standards of the user’s choice. (3) B-MIS normal-ization does not assume a uniform matrix effect among samplesand minimizes obscuring variation induced by sample-specificmatrix-induced ionization. (4) B-MIS normalization isindependent of and can be paired with other samplenormalization pre- or postacquisition. (5) B-MIS normal-ization is simple to apply, intuitive to interpret, and opensource.To demonstrate both the need for this method and its

functionality, we used two sample sets: marine environmentalsamples and laboratory cultures of marine phytoplankton. Weanalyzed these samples on two mass spectrometers: a WatersXevo TQ-S triple quadrupole, for targeted analyses, and aThermo QExactive HF, for untargeted analyses, using bothreversed phase (RP) and hydrophilic interaction liquidchromatography (HILIC). A comparative analysis of theseplatforms revealed the strengths of each and highlighted theneed for careful normalization and application of B-MISnormalization regardless of platform.

■ EXPERIMENTAL SECTION

The workflow for sample analysis and data processing is shownin Scheme 1. Additional details on sample preparation,chromatography, mass spectrometry, and quality controlparameters are described in the Supporting Information.Sample Preparation. We evaluated two sets of complex

samples in triplicate: cultures of one species of diatom grownunder four experimental conditions (salinity of 32 g/kg at both−1 and 4 °C, salinity of 41 g/kg at both −1 and 4 °C), and anatural marine microbial community at four depths in theNorth Pacific subtropical gyre (24° 33.284′ N, 156° 19.790′W; samples collected at 15, 45, 75, and 125 m, all withsalinities of 35 g/kg). Polar and nonpolar metabolites wereextracted using a modified Bligh−Dyer extraction25,26 using1:1 methanol/water (aqueous phase) and dichloromethane(organic phase). Methodological blanks were extracted andanalyzed along with each sample set. Isotope-labeled internalstandards were added either before or after the extraction to allsamples, blanks, and pooled samples (Table S1). To evaluatethe effect of obscuring variation due to different matrixstrengths and analytical drift, pooled samples were run at bothfull and half concentration (diluted with water) at least threetimes throughout a sample set.Liquid Chromatography. As in previous metabolomic

studies,2 we paired RP with HILIC (with separate injections)to maximize the number of compounds we were able to detectsince many compounds that eluted early on the RP columnwere better retained and provided higher quality peaks on theHILIC column (Figures S1 and S2). All chromatographicseparations were carried out on a Waters Acquity I-ClassUPLC (Waters Corporation, Milford, MA). For targetedanalysis, we monitored 105 analytes with RP and 126 analyteswith HILIC, according to which column demonstrated betterretention; 11 analytes were monitored on both columns tocompare the performance of RP and HILIC chromatographyin complex matrices (Table S2).Mass Spectrometry. For targeted metabolomics, we used

a Waters Xevo TQ-S triple quadrupole (TQS) with electro-spray ionization (ESI) in selected reaction monitoring mode(SRM) with polarity switching. SRM conditions for eachcompound (collision energy, cone voltage, precursor, andproduct ions, Table S2) were optimized by infusion of each

metabolite standard. For most metabolites, two SRMtransitions were selected based on maximum peak areas.We used a Thermo QExactive HF (QE) with ESI for

untargeted analyses. For HILIC, a full scan method withpolarity switching was used. For RP, positive ionization modewas chosen over positive/negative switching because few of thecompounds in our targeted method ionized in the negativemode and using a single mode enabled higher resolution.

Data Processing. For targeted analysis, we integratedpeaks using Skyline for small molecules.27 After integration, wepassed the data through an in-house quality control (QC) filterto ensure proper metabolite identification as described in theSupporting Information.For untargeted metabolomics data, we converted

Thermo.RAW files to .mzxml using MSConvert28 andprocessed each data set using XCMS29−31 (using parametersobtained via an Isotopologue Parameter Optimization32),yielding retention-time corrected mass features. We processedboth data sets in separate XCMS runs according to injectionand polarity (RP-aqueous, RP-organic, HILIC-aqueous (pos-itive polarity), and HILIC-aqueous (negative polarity)). Datawere quality controlled, as described in the SupportingInformation, prior to normalization.

B-MIS Normalization. The best normalization for eachanalyte was determined using repeat injections of a qualitycontrol sample to search for an isotope-labeled internalstandard whose obscuring variation matched the observedobscuring variation of the analyte, as shown in Scheme 1. Forour sample sets, multiple injections of the pooled sample at fulland half strength were used in order to capture the obscuringvariation due to variable matrix strength. After correcting fordilution, if the relative standard deviation (RSD) of ananalyte’s peak area in these repeat injections was less than 10%,the raw data were used. Otherwise, each analyte-internalstandard pair was evaluated for normalization. An internalstandard was deemed an acceptable match for an analyte if theanalyte’s RSD improved by 40% over the raw RSD, asdiscussed in the following section. If multiple internalstandards were acceptable for normalization of a single analyte,the one that minimized the RSD was selected. When a B-MISwas chosen, the peak area was divided by the ratio of the peakarea of the B-MIS to the average peak area of the B-MIS acrossthe whole sample set, resulting in an adjusted peak area for theanalyte. This adjusted peak area served as a relativequantification across samples in the sample set. If the goal ofan analysis is absolute quantification and standards areavailable, analytes can be quantified via standard addition ina subset of samples and then applied across the sample setusing the relative quantification of the B-MIS adjusted peaks.For internal standards with only one or two labeled atoms in

the compound, we examined the methodological blank toensure the signal was comparable to that in the samples,indicating that there was not significant contribution fromnaturally occurring isotopologues. For future users, we suggestchoosing isotope-labeled internal standards with at least a +2amu label to avoid any naturally occurring isotopologuecontamination. In the sample sets analyzed here, the isotope-labeled internal standards were significantly more abundantthat naturally occurring isotopologues, so contamination wasnegligible.

Analytical Chemistry Article

DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

C

■ RESULTS AND DISCUSSIONEvaluation of Data Quality and Instrument Perform-

ance. As in other studies,33−35 we found the HILIC columnproduced more variable chromatography than the RP column(higher median RSD of peak area, Figure 1A, see the exampleof choline in Figure S1, full results in Table S3). Despite lowerchromatographic reproducibility, compounds analyzed on theHILIC column generally exhibited less ion suppression thanthose analyzed on the RP column. RP chromatographydisplayed a bimodal distribution of response factor ratios(Figure 1B), demonstrating significant ion suppression formany analytes (response factor ratio < 1). Ion suppression wasespecially dramatic for polar compounds that elute early in RPbut were retained on the HILIC column (Figures S1 and S2).The range of ion suppression and RSD for a given LC-MSconfiguration indicated that careful selection of a normalizationtechnique was necessary to account for obscuring variation.This is particularly important when comparing samples incomplex matrices such as the environmental matrix highlightedin Figure 1. Furthermore, the wide range in response factorratios and RSD of peak areas for different analytes in the samematrix demonstrated that normalizing all compounds to asingle internal standard18,20 would likely introduce moreobscuring variation than it removed.Evaluating B-MIS Normalization. Most environmental

metabolomics analyses do not attempt to remove variabilityintroduced during analysis.5,6,8,36 B-MIS normalization wasdesigned to minimize obscuring variation in order to allowrobust comparisons within and between sample sets. Weevaluate B-MIS normalization by (a) demonstrating thatlowering the RSD of multiple injections of a pooled sample

improves data precision and arriving at a cutoff value at whichto apply normalization, (b) comparing the results from B-MISnormalization with the “gold standard” of correcting with anisotopologue for each analyte of interest, and (c) assessing theselection of internal standards used for normalization.To demonstrate the fundamental principle and evaluate the

effectiveness of B-MIS normalization, we applied the techniqueto the internal standards themselves. In other words, weevaluated how well the internal standards’ variabilities werematched by each of the other internal standards. Figure 2 is anexample of this application, where each point is the resultingRSD of D3-alanine in replicate injections of the pooled sample(x-axis) and across all samples in the environmental sample set(y-axis) after normalization to each possible internal standard;since we add the same amount of the internal standard to eachsample, both of these values should be near zero. Figure 2highlights that internal standards that reduced RSD of thepooled injections generally resulted in a reduced RSD acrosssamples (toward zero in both the x- and y-axes). However, inthis example case, there were two instances where a reductionin RSD of pooled sample injections increased the RSD of thesamples (marked on Figure 2 with an asterisk); thisintroduction of variability can be avoided by applyingnormalization only if the pooled RSD was improved beyonda cutoff value.Our goal was to maximize the number of compounds that

can be improved by B-MIS normalization while minimizing thelikelihood of introducing error. After assessing all internalstandards, we determined that requiring a 40% improvement ofpooled RSD (that is, (RSDfinal − RSDraw)/RSDraw) avoided themajority of situations where using an internal standard wouldresult in an increase of sample variability (Figures 2 and S3 andTable S6). Therefore, we chose this cutoff as the criteria toapply B-MIS normalization (see Scheme 1). Additionally, wedid not normalize any compounds or mass features that had araw pooled RSD of below 10% because the chances of

Figure 1. Box-and-whisker plots of (A) relative standard deviation(RSD) of peak areas determined by repeated injections of standardsspiked into our environmental sample matrix and (B) response factorratio on four LC−MS configurations, with individual dotsrepresenting single compounds (expanded horizontally for visual-ization of data density, n = 101 analytes in RP, 110 in HILIC).Response factor ratio is the ratio of the intensity of a standard injectedin environmental matrix (less the ambient matrix signal, if applicable)to the intensity of the standard injected in water. Response factor ratio< 1 indicates ion suppression, response factor ratio > 1 indicates ionenhancement.

Figure 2. RSD of D3-alanine in replicate injections of the pooledsample (x-axis) and across all samples in the environmental sample set(y-axis) after normalization to all possible internal standards. Theblack point represents the RSD of raw peak areas. The point at theorigin is D3-alanine normalized to itself. Filled points show thenormalization results with an acceptable matched internal standardaccording to our cutoff of 40% RSD improvement (left of dottedline); open points show normalization results using unacceptableinternal standards (right of dotted line, Scheme 1). Results from otherinternal standards are in Figure S3.

Analytical Chemistry Article

DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

D

significant improvement are low. Before applying B-MISnormalization, we suggest users perform a similar sensitivityanalysis to determine an appropriate cutoff using arepresentative sample set and internal standard suite. Toolsfor this analysis are available with B-MIS.The gold standard for absolute quantitation is an

isotopologue of each analyte, but isotope-labeled internalstandards are costly, consume duty time on the MS (intargeted analyses) and are not commercially available for manycompounds. In our targeted analyses, analytes with corre-sponding isotope-labeled internal standards were normalizedto those internal standards (see Scheme 1). If this was notprescribed, the algorithm occasionally selected a differentinternal standard, though the lowered RSD was often similar tothat of the prescribed B-MIS. For example, in the environ-mental data set, the amino acid valine could have selected theD7-proline internal standard based on the RSD reduction (rawpeak area RSD was 30.8%, which was minimized to 3.5% withnormalization to D7-proline). However, the “gold standard”internal standard, D8-valine, reduced the RSD to a similar4.8%, indicating proline and valine experienced nearly identicalobscuring variation in this matrix. For the other analytes thathave an isotope-labeled internal standard, the majorityincluded their corresponding internal standard as an acceptableB-MIS (>60% in both data sets, Table S5). Most cases wherethe corresponding internal standard was not consideredacceptable were due to the stringency of the 40% improvementcutoff. Nearly all of the analytes with an isotope-labeledinternal standard were less variable after normalization to theircorresponding internal standard, with one exception in eachdata set (pyridoxal in the environmental data set, cysteic acidin the culture data set; these exceptions were both improved bynormalization in the other data sets).The success of B-MIS normalization relies heavily on having

an appropriate suite of internal standards. However, if a massfeature does not match well with any internal standard, B-MISnormalization defaults to using the raw data, which a user canchoose to include in subsequent analysis. Increasing thenumber and chemical variety of internal standards results infewer features whose obscuring variation are not improved byB-MIS normalization (Figure S4). In our two data sets, the B-MIS for the targeted analytes were typically close in retentiontime (in RP, Figure S5) or similar in chemistry (in HILIC).For instance, amino acids generally matched with an isotope-labeled amino acid B-MIS while sulfonates generally matchedwith a sulfonate B-MIS (Table S5). Figure S4 shows that theinternal standards used in HILIC chromatography covered awide enough range of chemical diversity to effectively reduceobscuring variation in the majority of analytes detected underthese conditions.The specific choice of internal standards used in B-MIS

normalization is not prescribed and can vary from user to user;we suggest a selection of internal standards that spansretention time, m/z, and functional groups of interest. Beforeanalyzing a full sample set, users should determine theproportion of mass features which select each internal standardas “acceptable” and “best” matches by running a pooled orrepresentative sample at different dilutions. This will revealwhich, if any, internal standards are matching with a largenumber of mass features and may indicate that more internalstandards with similar chemical diversity could be beneficial(see Table S4). For example, in our untargeted diatom culturedata 13C3−vitamin B1 (thiamine) was selected as acceptable for

nearly 10% of mass features detected in the aqueous fractionanalyzed with RP chromatography and it was selected as the B-MIS for all of these features. It is likely that normalization ofthese compounds could improve if we included more internalstandards similar to 13C3-vitamin B1. Some internal standardswere infrequently selected as appropriate (such as 15N-isoleucine), while others were appropriate for many but onlythe best for a few (such as D7-proline), indicating sufficientcoverage to effectively normalize compounds that experiencedsimilar obscuring variation.

Application of B-MIS Normalization. B-MIS normal-ization only adjusts areas of metabolite peaks with highobscuring variation if normalization improves the technicalreproducibility above a specified cutoff (in our case, a 40%improvement over the raw technical reproducibility); there-fore, a subset of analytes are not normalized. Figure 3 shows an

example of two analytes (taurine and tyrosine) detected in thepooled diatom-culture sample whose RSD were improved byB-MIS normalization. For the sample sets analyzed here, 54%(diatom samples) and 79% (environmental samples) ofanalytes in the targeted data selected an internal standard fornormalization (Table S5 and Figure S5).For each compound, the RSD of the pooled injections

depends on the sample matrix, though the sample sets showeda similar range of RSD after B-MIS normalization (Figure S6).Of the 25 targeted analytes that were measured in both samplesets and found acceptable internal standards for normalization,20 shared acceptable internal standards, while five analytes didnot (Table S5 and Figure S6). The compounds which did notimprove with B-MIS normalization were not always consistentfrom environmental to culture sample sets. For example, therewere 18 compounds that did not pick a B-MIS in the targetedenvironmental sample set. In the culture sample set, six ofthese compounds selected a B-MIS, four did not select a B-MIS, and the remaining eight were not detected (Table S5).These differences between sample sets demonstrated that,while fundamental chemistry has some predictive capacity to

Figure 3. Example of B-MIS normalization for taurine and tyrosine inthe diatom sample set. (A) Areas of D4-taurine, taurine, and tyrosinein pooled samples after adjusting for dilution, in order of injection.(B) After normalization, areas are adjusted, with the B-MISnormalization reducing the RSD of taurine from 19% to 5%, whilethe RSD tyrosine decreases from 27% to 12% (black triangles).Normalizing to an unacceptable internal standard (13C2-sulfoaceticacid, gray triangles) did not reduce the RSD of tyrosine as effectively.

Analytical Chemistry Article

DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

E

describe obscuring variation, the complexity of the samplematrix makes definitive prediction unrealistic.In untargeted metabolomics, it is common to remove any

mass features that do not exhibit reliable data.37 One way thisis achieved is by calculating the RSD of each mass feature in amultiply injected pooled sample and removing features thathave a RSD greater than 20%.13,37 In the case of ouruntargeted data, B-MIS normalization increased the number ofmass features that passed this RSD filter from 3836 to 5028 (inthe diatom samples) and from 5568 to 9372 (in theenvironmental samples), preserving many high-quality peaksthat would otherwise have been discarded (Figure S7). In ourenvironmental untargeted analysis, 54% and 53% of the massfeatures detected on the HILIC and RP columns, respectively,matched with an acceptable internal standard and showeddecreased variability with a B-MIS (Figure S4). In this data set,a third of the most abundant and significantly different massfeatures would have been excluded if this RSD filtering hadbeen performed prior to B-MIS normalization (Figure 4). Atotal of 7 of the 10 mass features most enriched in the samplecollected at 125 m compared to the sample collected at 15 mwould have been excluded without B-MIS normalizationbefore filtering (Figure 4). This comparison highlights thatmeaningful biological interpretation of the differences betweensamples hinges on proper normalization. We thereforerecommend applying B-MIS normalization before filteringout mass features based on RSD.

■ CONCLUSIONS

Published environmental metabolomics analyses have notattempted to minimize the obscuring variation inherent to LC-MS-based data acquisition. In both targeted and untargetedenvironmental samples, we demonstrate the need for carefulnormalization due to matrix-induced ion suppression andanalytical drift. This manuscript introduces, evaluates, andapplies a batch-specific, simple, and customizable normal-ization process: Best-Matched Internal Standard (B-MIS)normalization. By applying B-MIS normalization, we are ableto retain many more LC-MS mass features during untargetedRSD filtering. For future users, we have made the tools tocustomize, apply, and evaluate B-MIS normalization freelyavailable at https://github.com/IngallsLabUW/B-MIS-normalization.

■ ASSOCIATED CONTENT*S Supporting InformationThe Supporting Information is available free of charge on theACS Publications website at DOI: 10.1021/acs.anal-chem.7b04400.

• Supplementary Methods; Table S1; Figures S1−S7; andcaptions for Tables S2−S6 PDF.

• Tables S2−S6 (XLSX)

■ AUTHOR INFORMATIONCorresponding Author*E-mail: aingalls@uw.edu. Phone: (206) 221-6748.ORCIDKatherine R. Heal: 0000-0002-4504-1039Author Contributions‡These authors contributed equally to this work, orderdetermined by eyeglass prescription strength.NotesThe authors declare no competing financial interest.

■ ACKNOWLEDGMENTSThe authors acknowledge L. Hmelo, J. Collins, R. Lundeen, R.Lionheart, A. Weid, and N. Kellogg for assistance with labwork, data analysis, or manuscript writing; J. Young and H.Dawson for providing the diatom cultures; B. Durham forproviding several standards; and the crew and scientific partyof the R/V Kilo Moana during HOE-Legacy 2A. Thanks tofour anonymous reviewers whose comments greatly improvedthis manuscript. This work was supported by grants from theSimons Foundation (LS Award ID: 385428, A.E.I.; SCOPEAward ID 329108, A.E.I.), NSF OCE-1228770 and OCE-1205232 to A.E.I., NSF GRFP to K.R.H., NSF IGERTProgram on Ocean Change to A.K.B.

■ REFERENCES(1) Gika, H. G.; Theodoridis, G. A.; Plumb, R. S.; Wilson, I. D. J.Pharm. Biomed. Anal. 2014, 87, 12−25.(2) Hounoum, B. M.; Blasco, H.; Emond, P.; Mavel, S. TrAC, TrendsAnal. Chem. 2016, 75, 118−128.(3) Bundy, J. G.; Davey, M. P.; Viant, M. R. Metabolomics 2009, 5,3−21.(4) Sysi-Aho, M.; Katajamaa, M.; Yetukuri, L.; Oresic, M. BMCBioinf. 2007, 8, 93.(5) Kido Soule, M. C.; Longnecker, K.; Johnson, W. M.; Kujawinski,E. B. Mar. Chem. 2015, 177, 374−387.

Figure 4. Log2-fold change between surface (15 m) and deep (125 m) environmental samples for the 500 most abundant mass features in eachsample type that were significantly different between depths (false discovery rate38 adjusted p-values < 0.05, Student’s t-test). A positive fold changeindicates enrichment in the 125 m samples. Many mass features (292 of the 883 shown here) may not have been considered reliable peaks (RSD >20%) due to high obscuring variation which was minimized through B-MIS normalization; these mass features are highlighted in red.

Analytical Chemistry Article

DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

F

(6) Llewellyn, C. A.; Sommer, U.; Dupont, C. L.; Allen, A. E.; Viant,M. R. Prog. Oceanogr. 2015, 137, 421−433.(7) Johnson, W. M.; Kido Soule, M. C.; Kujawinski, E. B. ISME J.2016, 10, 2304−2316.(8) Barofsky, A.; Vidoudez, C.; Pohnert, G. Limnol. Oceanogr.:Methods 2009, 7, 382−390.(9) Swenson, T. L.; Jenkins, S.; Bowen, B. P.; Northen, T. R. SoilBiol. Biochem. 2015, 80, 189−198.(10) Baran, R.; Bowen, B. P.; Bouskill, N. J.; Brodie, E. L.; Yannone,S. M.; Northen, T. R. Anal. Chem. 2010, 82, 9034−9042.(11) Paul, C.; Mausz, M. A.; Pohnert, G. Metabolomics 2013, 9,349−359.(12) Johnson, W. M.; Kido Soule, M. C.; Kujawinski, E. B. Limnol.Oceanogr.: Methods 2017, 15, 417−428.(13) Dunn, W. B.; Broadhurst, D.; Begley, P.; Zelena, E.; Francis-McIntyre, S.; Anderson, N.; Brown, M.; Knowles, J. D.; Halsall, A.;Haselden, J. N.; Nicholls, A. W.; Wilson, I. D.; Keil, D. B.; Goodacre,R.; Consortium, T. H. S. M. Nat. Protoc. 2011, 6, 1060−1083.(14) Calderon-Santiago, M.; Lopez-Bascon, M. A.; Peralbo-Molina,A.; Priego-Capote, F. Talanta 2017, 174, 29−37.(15) Shen, X.; Gong, X.; Cai, Y.; Guo, Y.; Tu, J.; Li, H.; Zhang, T.;Wang, J.; Xue, F.; Zhu, Z.-J. Metabolomics 2016, 12, 1−12.(16) Gagnebin, Y.; Tonoli, D.; Lescuyer, P.; Ponte, B.; de Seigneux,S.; Martin, P.-Y.; Schappler, J.; Boccard, J.; Rudaz, S. Anal. Chim. Acta2017, 955, 27−35.(17) Wu, Y.; Li, L. J. Chromatogr. A 2016, 1430, 80−95.(18) Bromke, M. A.; Sabir, J. S.; Alfassi, F. A.; Hajarah, N. H.; Kabli,S. A.; Al-Malki, A. L.; Ashworth, M. P.; Meret, M.; Jansen, R. K.;Willmitzer, L. PLoS One 2015, 10, e0138965.(19) Fei, F.; Bowdish, D. M.; McCarry, B. E. Anal. Bioanal. Chem.2014, 406, 3723−3733.(20) Bijlsma, S.; Bobeldijk, I.; Verheij, E. R.; Ramaker, R.; Kochhar,S.; Macdonald, I. A.; Van Ommen, B.; Smilde, A. K. Anal. Chem.2006, 78, 567−574.(21) Heal, K. R.; Carlson, L. T.; Devol, A. H.; Armbrust, E.; Moffett,J. W.; Stahl, D. A.; Ingalls, A. E. Rapid Commun. Mass Spectrom. 2014,28, 2398−2404.(22) Heal, K. R.; Qin, W.; Ribalet, F.; Bertagnolli, A. D.; Coyote-Maestas, W.; Hmelo, L. R.; Moffett, J. W.; Devol, A. H.; Armbrust, E.V.; Stahl, D. A.; Ingalls, A. E. Proc. Natl. Acad. Sci. U. S. A. 2017, 114,364−369.(23) Bueschl, C.; Krska, R.; Kluger, B.; Schuhmacher, R. Anal.Bioanal. Chem. 2013, 405, 27−33.(24) Katajamaa, M.; Oresic, M. BMC Bioinf. 2005, 6, 179.(25) Bligh, E. G.; Dyer, W. J. Can. J. Biochem. Physiol. 1959, 37,911−917.(26) Canelas, A.; ten Pierick, A.; Ras, C.; Seifar, R.; van Dam, J.; vanGulik, W.; Heijnen, J. Anal. Chem. 2009, 81, 7379−7389.(27) MacLean, B.; Tomazela, D. M.; Shulman, N.; Chambers, M.;Finney, G. L.; Frewen, B.; Kern, R.; Tabb, D. L.; Liebler, D. C.;MacCoss, M. J. Bioinformatics 2010, 26, 966−968.(28) Chambers, M. C.; et al. Nat. Biotechnol. 2012, 30, 918−920.(29) Smith, C. A.; Want, E. J.; O’Maille, G.; Abagyan, R.; Siuzdak, G.Anal. Chem. 2006, 78, 779−787.(30) Tautenhahn, R.; Bottcher, C.; Neumann, S. BMC Bioinf. 2008,9, 504.(31) Benton, H. P.; Want, E. J.; Ebbels, T. M. Bioinformatics 2010,26, 2488−2489.(32) Libiseller, G.; Dvorzak, M.; Kleb, U.; Gander, E.; Eisenberg, T.;Madeo, F.; Neumann, S.; Trausinger, G.; Sinner, F.; Pieber, T.;Magnes, C. BMC Bioinf. 2015, 16, 118.(33) Contrepois, K.; Jiang, L.; Snyder, M.Mol. Cell. Proteomics 2015,14, 1684−1695.(34) Cubbon, S.; Antonio, C.; Wilson, J.; Thomas-Oates, J. MassSpectrom. Rev. 2010, 29, 671−684.(35) Spagou, K.; Tsoukali, H.; Raikos, N.; Gika, H.; Wilson, I. D.;Theodoridis, G. J. Sep. Sci. 2010, 33, 716−727.

(36) Poulson-Ellestad, K. L.; Jones, C. M.; Roy, J.; Viant, M. R.;Fernandez, F. M.; Kubanek, J.; Nunn, B. L. Proc. Natl. Acad. Sci. U. S.A. 2014, 111, 9009−9014.(37) Vinaixa, M.; Samino, S.; Saez, I.; Duran, J.; Guinovart, J. J.;Yanes, O. Metabolites 2012, 2, 775−795.(38) Benjamini, Y.; Hochberg, Y. J. R. Stat. Soc. Ser. B 1995, 289−300.

Analytical Chemistry Article

DOI: 10.1021/acs.analchem.7b04400Anal. Chem. XXXX, XXX, XXX−XXX

G

View publication statsView publication stats